Basel Action Network’s E-Waste Report Turns AI’s Hardware Refresh Into a Numbers Fight

AI infrastructure could generate between 395 million and 617 million tonnes of electronic waste between 2025 and 2050, enough to fill 15 million to 23 million 40-foot shipping containers. Laid end to end, 20 million of them, a representative figure within that range, would stretch roughly 244,000 kilometers, about six times around the Earth. That is the headline number from a new report by the Basel Action Network (BAN), a Seattle-based nonprofit that monitors international hazardous waste trade under the 1989 Basel Convention, published September 15 and since covered by outlets including The Register. But the report’s more substantive argument is narrower than its headline: nearly every prior estimate of AI’s waste footprint counted only a small slice of what actually fills a data center, and independent analysts who reviewed the new numbers say the report’s own most dramatic claim, that this dwarfs those estimates by 40 to 60 times, rests on an assumption that is already being contradicted by how hyperscalers account for their own hardware.

What Prior Estimates Left Out

The report, titled How Big Is the AI Waste Wave?, is the first of a planned four-part series written by BAN founder and chief of strategic direction Jim Puckett. Its central claim is about scope, not just scale. Servers and the GPUs and other accelerators inside them, the equipment most existing e-waste studies count, add up to only about 13 percent of a data center’s electromechanical infrastructure by weight, according to the report. The remaining 87 percent, cooling systems, power distribution gear, backup power supplies, and networking equipment, “has never appeared in any AI e-waste projection of which we are aware,” the report states.

To size that overlooked majority, BAN modeled a reference 100-megawatt AI data center and broke its hardware mass into five categories: cooling, power distribution, backup power, servers and accelerators, and networking. Frank Dickson, principal analyst at Dickson Research, who reviewed the methodology for Network World, said the report weights those categories at cooling 35 percent, power distribution 34 percent, backup power 15 percent, servers and accelerators 13 percent, and networking 3 percent, together totaling roughly 70,000 metric tonnes per gigawatt of capacity, or about 7,000 tonnes for the 100-megawatt reference facility. BAN cross-checked that per-gigawatt figure against the World Economic Forum’s mineral-intensity data and Microsoft’s own disclosed copper consumption at a Chicago facility, Dickson said. “That’s a reasonably rigorous way to build a per-gigawatt hardware mass estimate and I don’t have a strong basis to dispute the 62,000-to-77,000-tonne-per-gigawatt range it lands on.” A single 100-megawatt facility, the report notes, contains about 2,700 tonnes of copper in its cabling and busbars alone, mostly because upgrading from 5 to 15 kW server racks to the 50 to 140 kW racks AI workloads require means ripping out and replacing much of the power infrastructure, not just swapping chips.

“To date the environmental debate around AI has focused on electricity, carbon and water while largely overlooking what happens to the hardware itself,” Puckett said. “If companies and governments do not begin planning for this new waste tsunami, today’s AI buildout could become an even more cataclysmic toxic waste crisis than we are already experiencing.”

The Math Behind Six Trips Around the Earth

Puckett layers that per-facility baseline onto a global buildout forecast: roughly 7 trillion dollars in AI infrastructure spending worldwide between 2025 and 2030, enough to add 219 gigawatts of data center capacity, more than the power needed for every home in the United States, he estimates. The equipment inside those facilities does not last long under BAN’s model. GPUs and other accelerators are assumed to turn over roughly every two and a half years, compared with 5 to 7 years historically for general-purpose servers. Networking gear is replaced every 3 to 4 years, backup power and cooling systems every 5 years, and power distribution equipment every 8 years, all considerably faster than the multi-decade lifespans traditional infrastructure has typically had.

Run that replacement math against BAN’s growth curve, an 8.8 percent compound annual capacity growth rate sustained across all 26 years from 2025 to 2050, and the total lands at 395 million to 617 million tonnes of AI-driven electronic equipment retired over that period. “The cumulative AI-driven electronic equipment being retired between 2025 and 2050 (395 to 617 million tonnes) would fill 15 million to 23 million 40-foot shipping containers,” Puckett wrote, quoted in the trade publication Resource Recycling. “Placed end to end, they would circle the Earth about six times.” Narrowed to a single year, the report projects that by 2030, AI-driven equipment retirement will run roughly 40 to 60 times higher than the most widely cited academic projection, a gap BAN attributes almost entirely to the fact that prior studies counted only servers and GPUs.

Puckett’s model also includes a second, less obvious waste stream he calls AI waste contagion: computers, phones, and telecom equipment outside data centers that become obsolete faster because they cannot keep pace with AI-driven software demands. He estimates that contagion at 16 million to 30 million tonnes a year, larger than the roughly 15.5 million tonnes a year his model projects will come out of data centers themselves. Zoomed out further, the report projects total global e-waste generation from all sources, not AI alone, reaching 196 million to 211 million tonnes a year by 2050, more than triple today’s roughly 67 million tonnes annually. Of that 2050 total, BAN attributes 31 million to 46 million tonnes a year specifically to AI.

The Number Analysts Don’t Buy

Analysts who reviewed the report for Network World generally accepted BAN’s per-gigawatt hardware mass estimate but pushed back hard on the multiplier built on top of it. “The headline number, 40 to 60 times prior estimates, is going to get all the attention. It is the least defensible part of this report,” Dickson said, adding that the rigor of the mass calculation is precisely what makes the shakier assumptions stand out. “It leans on assumptions doing a lot of work: an 8.8 percent compound annual growth rate in data center capacity sustained for 26 straight years, and a 2.5-year retirement cycle for accelerators specifically. That last one is problematic because it’s now a live industry debate.”

Dickson’s specific objection is grounded in public accounting disclosures. Every major hyperscaler extended its own useful-life assumptions for servers between 2022 and 2025, he said: Microsoft moved from 4 to 6 years, Alphabet from 4.5 to 6, Meta to 5.5, and Oracle from 5 to 6. Those changes were based on the argument that a chip’s working life does not end when it leaves a frontier training cluster, but cascades down into inference and lower-intensity batch work for years afterward. “BAN’s own report acknowledges this research exists and calls it untested for AI accelerators specifically, then keeps the 2.5 year figure anyway,” Dickson said. “It is the report’s weakest link. The honest read is that BAN’s near-term, 2030-era numbers are probably overstated.” He did credit one part of BAN’s underlying logic, however: “AI-driven power density is compressing replacement cycles for cooling and power-distribution gear faster than most capital planning models have caught up to.”

Independent technology consultant Steven Eric Fisher raised a more fundamental objection to how the report defines waste in the first place. “Equipment being retired from a particular installation is not necessarily the same thing as equipment becoming waste,” he said. “The report assigns accelerators, servers, and racks a 2.5 year lifespan and connects that partly to Nvidia’s architectural release cadence. I don’t think product generation cadence can be used as a proxy for useful equipment life.” As evidence, Fisher pointed out that Nvidia introduced its V100 accelerator in 2017 and Google Cloud still lists V100 instances today, while the A100, introduced in 2020, remains an actively offered AWS platform in 2026. “That does not mean every hyperscaler operates hardware that long, but it demonstrates that a newer generation entering the market does not automatically make the previous generation economically useless,” he said. Fisher raised the same concern about the report’s other lifespan assumptions: “The report itself states that traditional power infrastructure can last 15 to 20 years. The five year cooling lifespan is also a BAN estimate. Once those assumptions are multiplied across hundreds of gigawatts of projected capacity, relatively small modeling choices can produce extremely large waste totals.”

Nidhi Luthra, an executive advisor at Acceligence, framed the report’s practical value differently than its headline suggests. “I think the most important thing about this report is not whether every long range tonnage estimate proves exact. It is that IT may be measuring the wrong thing,” she said. “The approximately 70,000 tonnes per gigawatt estimate is directionally plausible based on the methodology they lay out. But the 2050 projections are much more assumption-sensitive.” She pointed to a distinction that cuts across both sides of the lifespan debate: “The bigger executive issue is that AI may create economic obsolescence faster than physical obsolescence. Equipment can still work perfectly well and yet become commercially unattractive because the next generation requires different power density, cooling, networking, or rack architecture.” Her broader point: “AI has largely been discussed as a software, compute and energy story,” she said. “It is increasingly becoming a materials and lifecycle management story as well.”

Retired Doesn’t Always Mean Trash

Puckett himself does not dismiss the resale-and-reuse counterargument his critics raise. He acknowledges that equipment lifespans exceeding current expectations, and modular servers that allow part upgrades rather than total replacement, could both push his estimates down, though he argues current industry trends point the other way. He also flags the refurbishment and resale market as the report’s biggest open variable. “Many in the ITAD industry are viewing the AI horserace as a gold mine for business, the most lucrative of which will be refurbishing and reselling the surplus hardware following the buildout boom and likely rapid refresh cycles,” he wrote, referring to the IT asset disposition sector that resells and recycles retired enterprise hardware. “Surely there will be a massive demand for fast but not fastest equipment in developing countries, in lower-tier data centers and inference workloads. The question is a valid and open one: Will reuse save us from an e-waste tsunami? Surely the hyperscalers are planning to ensure such a future.”

That question remains unanswered in a literal sense: according to the report, no hyperscaler, government, or international body has published a plan for handling AI-driven e-waste at the scale it projects, and BAN argues there is insufficient infrastructure to safely process even today’s e-waste volumes, let alone a coming surge. Parts 2 through 4 of the series, still to come, are meant to fill in the picture the first paper leaves open: quantifying how much reuse and refurbishment could realistically offset the waste, assessing the toxicity of what gets left behind, including contamination from PFAS, and proposing ways to mitigate the risk.

The report lands amid a broader pattern of AI infrastructure drawing scrutiny for costs that were not part of the original pitch. Sxz.io has covered former EPA officials warning that data center deregulation could cost 1,300 lives a year from the power-plant pollution needed to fuel the buildout, and China’s rare earth export pause turning into a supply deadline for the materials that go into building that same hardware in the first place. BAN’s report argues the other end of that materials pipeline, what happens when the hardware comes back out, has gotten even less attention than either.